5 Things Uber Just Got Right With AI

Uber’s engineering leadership posted a long thread on X last week, and the headline numbers are the part everyone screenshotted: 99% of their engineers on AI tools, 70%+ of pull requests now coming from agents, 2,500+ agent skills in production. Impressive. Also not the interesting part. The interesting part is how they pushed this past engineering, into Finance, Legal, Ops, Marketing, Support, HR, and Procurement, using two-week “Agentic Pods.”

Full disclosure before anyone gets the wrong idea: we don’t work with Uber. This isn’t a case study of ours, and there’s no relationship to imply. It’s just the clearest public example we’ve seen of a big company doing what we (and a lot of operators) have been saying for months. Plenty of people are converging on this independently. Uber just published the receipts.

Don’t automate the task. Rebuild the workflow.

We’ve hammered this one. In “Cuban’s right. The AI bill is for the workflow you never rebuilt” and “Bolt a motor to a line shaft and you’ve still got a line shaft”, the point was the same: bolt AI onto a broken process and you’ve got a faster broken process.

Uber, in their words: “The biggest wins rarely come from automating one task. Once you redesign the workflow around AI, you eliminate handoffs, remove unnecessary approvals… The workflow becomes the unit of automation — not the individual task.”

The result: capital allocation across 150 cities went from 15 hours to 30 minutes. Financial pacing reports went from 2 days to 10 minutes.

For a 30-person shop, the line-shaft is your quote-to-invoice chain. If a job passes through four inboxes and two approvals before anyone bills it, automating the “type the invoice” step saves you six minutes. Redesign the whole chain and you cut days off the cash cycle. Say your average job is $4,000 and you’re collecting 40 days out. Pull that to 20 (an example, your numbers will vary) and you’ve freed up real working capital you can actually spend.

Process diagrams lie. Go watch the work.

This is the whole thesis of “Why Process Beats Raw Capability”. Your bottleneck is understanding how the work actually happens. The model is already good enough.

Uber agrees, hard: “You can’t automate them effectively by looking at process diagrams or documentation. You have to understand how the work actually gets done.” Days 1 and 2 of every pod are spent shadowing the expert and observing every step. Their closing line: “You discover them by sitting next to the people doing the work… building with them, not for them.”

The result: engineers dropped into domains they’d never touched found opportunities “hiding in plain sight,” the stuff no flowchart captures because the person doing it stopped writing it down years ago.

Your front-desk person has fourteen little workarounds that live nowhere. The scheduling exception for the customer who only pays by check. The vendor you always call instead of email. Sit next to them for two days with a notepad. That’s the entire discovery budget, and it’s free.

Stop evaluating. Start building.

We’ve written this one twice, loudly: “You can just install Claude Code tonight” and “You can just decide to be AI-native”. The permission you’re waiting for doesn’t exist. You can start today.

Uber gave every pod exactly two weeks. Day 10 is labeled “Ship.” Something works in production by then, or the pod failed. No interim deliverable to hide behind.

The result: 16 Agentic Pods across 16 business functions in the past two months.

Compare that to how most SMBs approach this: a six-month evaluation, a steering committee, a vendor bake-off, and a pilot that never leaves the parking lot. Uber has 30,000 engineers and moves in fourteen-day cycles. A 30-person company deciding it needs a quarter to “assess readiness” has the process backwards. You can run the two-week version next week.

The best opportunities cross departments.

“Your Job Was Never Typing” was about this exact seam. The value lives in the ugly handoffs between roles nobody owns, the gaps that fall off the end of every job description.

Uber: “The most impactful agent skills cut across teams, orgs, functions, tools, and systems.”

The result: marketing web QA dropped from 2 weeks to 50 minutes. Support went from 9,000 manual workflows to self-service automation.

At your scale the cross-department seam is where sales meets ops meets accounting. The deal that closes in the CRM but takes a week to become a scheduled job with a PO. Nobody owns that gap because it’s between three people’s job descriptions. That gap is where your best week of work is hiding.

Redesigning the work kills the vendor tax.

We wrote “The SaaS Tax: How Enterprise Software Became Your Most Expensive Employee” about the subscriptions that add up to a full salary while nobody’s watching. Uber found the same lever. The savings lived in the workflow redesign.

Uber, on redesigning the workflow: it lets you “replace legacy tooling, reduce vendor spend.”

The result worth noting: the savings showed up when they rebuilt the workflow, and doing a task faster never touched the subscription. A tool retires when you redesign the process so hard that you no longer need what it did.

A 30-person shop is often renting eight to twelve tools at $15 to $80 a seat. Automate the task and you keep paying for all of them. Rebuild the workflow and two or three of those subscriptions stop making sense. That’s the line item that actually moves.

None of this is new to anyone who’s read us for a while. It’s the through-line we’ve published for months: rebuild the workflow, watch the real work, ship before you’re ready. Uber didn’t invent the idea. They just gave it a clean distillation and attached numbers big enough that people finally look up.

Here’s what stays with us. Uber has 30,000 engineers, 2,500 agent skills, and a platform team most companies would kill for. And their number-one lesson is still: sit next to the person doing the work for two days. That’s it. That’s the barrier to entry. A 30-person company can run that exact play next week: no steering committee, no readiness assessment, no permission slip. The expertise you need is already sitting at the front desk. Somebody just has to pull up a chair.